Microscaling (MX): definition · Chip Field Guide

Beginner

Giving each small group of numbers its own shared multiplier, so one huge number only affects its own group.

Novice

Block-scaled formats standardized by the Open Compute Project: every 32 consecutive values share one 8-bit power-of-two scale, and each value is stored in FP8, FP6, FP4 or INT8. MXFP4, for example, costs 4.25 bits per value.

Expert

OCP MX v1.0: block size $k = 32$ , scale $X$ in E8M0 ( $2^{-127}$ to $2^{127}$ , one NaN code), elements $P_i$ in E4M3/E5M2, E3M2/E2M3, E2M1 or INT8. $X = 2^{\lfloor \log_2 \max|v| \rfloor - e_{\max,\mathrm{elem}}}$ ; $P_i = \mathrm{round}(v_i/X)$ with saturation. A dot product computes $X_A X_B \sum P_A P_B$ , with internal precision implementation-defined.